Grapheme-to-phoneme conversion with neural networks: automatic morphological analysis

Abstract

The paper provides an alternative approach to one of the key steps of speech synthesis – grapheme-tophoneme conversion. The approach is based on automatic morphological analysis of out-of-vocabulary words into their consitutent morphemes. In constrast to a traditional method of morphological analysis based on finite state transducers (FST), we propose here a solution that makes use of neural networks. Experiments show that morphological analysis with the proposed neural network approach is significantly more effective (accuracy 93.8%) than morphological analysis with FST (accuracy 75%). The proposed approach allows to speed up the creation of new grapheme-to-phoneme models, as well as making it easier to build and maintain pronunciation dictionaries.